Users tend to use their own terms to search items in structured search systems such as restaurant searches (e.g. Yelp), but due to users' lack of understanding on internal vocabulary and structures, they often fail to adequately search, which leads to unsatisfying search results. In this case, search systems should assist users to use different terms for better search results. To address this issue, we develop a scheme to generate suggested queries, given a user query. At first in doing this, in this paper we propose a scheme to evaluate queries (i.e. user queries and/or suggested queries) based on two measures: 1) if the query will return a sufficient number of search results, 2) if the query will return search results of high quality. Furthermore, we present a learning model to choose among alternative candidate queries against a user query. Our experiments show the proposed method is feasible and scalable.
Quality of queries, query suggestion, learning to measure
The generic Internet search does not confront empty search results, such as web search, which always return millions of web pages (e.g. Google). However, due to users’ lack of understanding on internal vocabulary and structures, if they use their own terms to search items in controlled vocabulary structure-based search systems, such as restaurant search on Yelp, they may obtain empty, too many or unsatisfying results. In this case, the system should suggest better query terms to users. Although generic search engines (e.g. Google) may also have query suggestion schemes based on most users’ co-click frequency, they may not fully utilize controlled and structured domain vocabulary in specific areas (e.g. restaurant search), in which users will expect search results of a higher level than generic search systems provide. To assist users use better search terms, we develop a scheme to suggest alternative queries, given a user’s query. For doing this, we evaluate the quality of user’s query and/or suggested queries. Hence, we propose two measures: 1) whether the query will return a sufficient number of search results, 2) and/or whether the query will return search results of high quality. Based on the two measures, we build query suggestion and measurement models in one of restaurant
For restaurant search, the most common types of search term are
location and food. Figure 1 shows our query suggestion model
how to deal with both. According to different inputs, the model
suggests different options. First, the model checks the types of
search terms: Original query is: 1) single food; 2) single
location; 3) food with location. For case 1) and 2), the model has
four choices: relax to bigger scope terms (R), change to similar
terms (C), tighten (T) to smaller scope ones and do nothing (N);
for case 3), since there are four types of actions: N, R, C and T
and two types of terms, thus the model have options to
suggest, for instance, RC means relaxing food with changing
location. This scheme is scalable for more types of terms, e.g., n
types, options to suggest.
In previous work, knowledge-ba.sed recommender systems [
In controlled vocabulary structure-based searches, there are few researches about query measurement. Here, We evaluate queries based on two measures:
Quantity: The query should return a sufficient number of results. For controlled vocabulary structure-based search, the
number of returned results can be neither empty nor too much. Since the number of restaurants diversely distribute in the whole Korea, we can set neither a fixed number nor an interval to measure the quantity of results.
Quality: The query should return search results with high
quality. To define the high quality restaurant (HQ-rest), we
collected blogs from NAVER blogosphere in Korea
(http://blog.naver.com/) and ranked restaurants according to
power bloggers’ contribution [
Neither quantity nor quality can be measured by simple threshold or an interval. Therefore, we build a learning model to classify query into good or bad, with six criterions in terms of quantity and quality. The criterions are as follows:
The number of matched restaurants among top 10 in returned list: matched restaurant is defined as the restaurant, which satisfies user’s requirements. (E.g., user inputs “chicken” and try to find fried chicken restaurant, but the returned results include other types of chicken, e.g. chicken soup. In this case, we count the number of fried chicken restaurants as the number of matched restaurants.) The ranked position of first matched restaurant;
We make training data by evaluating 302 queries with human
effects. 76 are evaluated as good, the other are as bad. We build
learning model for several algorithms. The accuracies of
10-folds cross validation, precision, recall and F-score are
shown in Table 1. The average accuracy is more than 77%.
We draw the distribution of first matched position with number
of HQ-rest for each location level in Figure 2. It indicates:
#HQ-rest in high-level location is more than low-level location;
the first matched position ranges [
58.5% 71.4% 82.5% 79.3% 86.9%
97.6% 47.1% 94.1% 76.5% 91.3%
73.1% 56.7% 87.9% 77.8% 89%
In this paper, to assist users to search better results in restaurant searches, we propose a scheme to suggest alternative queries and present a learning model to measure queries in terms of quantity and quality. Query suggestion can be extended for adding more types of search terms. The query measurement experiments showed that our method is feasible and scalable. Our future work includes: applying probabilistic model to rest-food matrix (e.g. PLSA, LDA), accumulating more training set and applying heuristic, ranking methods to query measurement.
This research is supported by Mid-career Researcher Program through NRF grant funded by the MEST (2011- 0029729).